The Neuroscience of Intelligence

by Richard J Haier

  The structural image in Figure 3.4 shows gray matter, where neurons work, and the white matter fibers that link brain areas and carry information around the brain. Gray matter and white matter tissue have different water content, so they can be distinguished in these images. Note that structural images do not contain any functional information, so you cannot look at a structural MRI and tell if the person is awake or asleep, solving math problems, or even alive or dead. You can see tumors, strokes, and many kinds of brain damage. MRI can also be used to show brain function. Very rapid sequential images can show regional blood flow as a function of hemoglobin determinations and blood flow is an indirect measure of neuron activity. The more a brain area is active, the more blood flows to it. It is functional MRI (fMRI) that has been used widely in cognitive neuroscience to show brain activity during specific cognitive tasks.

  The basic MRI technique is quite versatile. By changing various parameters of the scanning sequence, for example, like the frequency of the radio wave pulses, different kinds of pictures can be made that emphasize different brain characteristics. As noted, the two main kinds of MRI are structural and functional. Structural MRI methods include the basic scan which shows gray and white matter in anatomical detail and other methods that maximize imaging of white matter fibers and tracts like spectroscopy (MRS) and diffusion tensor imaging (DTI). Such structural images are not affected by what the brain is doing during the scan. We’ll now review the early intelligence studies that used structural and functional MRI. We start with the basic structural MRI, because this was the first way MRI was applied to intelligence research.

  3.6 Basic Structural MRI Findings

  The first question about intelligence addressed by MRI had to do with whole brain size. Numerous previous studies had reported a positive correlation between brain size and intelligence test scores. The correlation typically was modest, but the main problem was that the measures of brain size were estimates based on indirect measures like head circumference (or in the 1800s the number of metal pellets required to fill a skull). MRI provided a much more exact measurement of brain size/volume in vivo, so it was not surprising to find confirmation of a positive correlation with intelligence test scores when accurate MRI-based measurements of brain size were used (Willerman et al., 1991). This is a straightforward finding that has been replicated many times. A comprehensive meta-analysis of this literature (37 studies, 1,530 subjects) reported an average correlation between whole brain size/volume and intelligence test scores to be about .33 overall (McDaniel, 2005), including adults and children. The correlation was higher in females (about .40 compared to .34 in males). In female adults and children, the correlations were .41 and .37, respectively. In male adults and children, the correlations were even more different at .38 and .22, respectively. These data essentially resolve the earlier debate and show definitively that bigger brains are modestly associated with higher intelligence.

  Of course, questions remain. Are the volumes of some specific brain areas more related to intelligence than other areas? What influences the development of brain size, and can the developmental mechanisms be accentuated? We’ll discuss the latter question in Chapter 5 about enhancing intelligence. The former question was addressed soon after structural MRIs were augmented with image-analysis methods that segmented or “parcellated” cortical and subcortical areas into regions of interest (ROIs). ROIs typically were derived either by applying a simple algorithm based on an arbitrary proportion of voxels thought to define a region or by human observers tracing ROIs on each image to the best of their ability using various brain landmarks. These early segmentation methods varied among research groups and all were rudimentary by today’s standards, but the results did indicate that size/volume of some areas was more related to intelligence than it was in other areas. One group (Andreasen et al., 1993), for example, reported small positive correlations between Full-Scale IQ (FSIQ) and volume of the temporal lobes, hippocampus, and cerebellum, and Flashman et al. (1997) further reported small correlations with Performance (non-verbal) IQ in frontal, temporal, and parietal lobes. None of these correlations exceeded whole brain/IQ correlations, but they hinted at the importance of regional analyses.

  3.7 Improved MRI Analyses Yield Consistent and Inconsistent Results

  By the time the next structural MRI studies of intelligence were reported, image analyses had improved spatial localization by replacing ROI segmentation with methods that quantified gray and white matter voxel by voxel (Ashburner & Friston, 1997, 2000) with spatial resolution of millimeters rather than lobes or parts of lobes. Software for the application of voxel-based morphometry (VBM) became available in about 1999 (statistical parametric mapping, SPM) and the field moved dramatically away from customized image analysis based on ROIs that differed in their boundaries among research groups to a more standardized approach. Typically, the results of voxel-based analyses were reported as spatial locations in the brain using a standard set of coordinates developed at the Montreal Neurological Institute (MNI coordinates) and the locations were described additionally using a standard nomenclature based on Brodmann areas (BAs) derived from early autopsy descriptions of different cellular organization among cortical regions (Brodmann, 1909). Textbox 3.1 describes the VBM method and includes an illustration of BAs. SPM is updated periodically with improvements and additional options for analyses.

  Textbox 3.1: Voxel-based morphometry

  One main method for analyzing MRI used on structural or functional images is called voxel-based morphometry, or VBM. There are three basic steps, as shown in Figure 3.5. First we start with an image like the MRI on the left. Next, mathematical algorithms determine the boundaries of gray and white matter tissue. Finally, values are calculated that reflect the amount of gray or white matter tissue in each voxel in the whole brain. Because there are millions of voxels in the whole-brain image, you get a very large data set. You can then correlate a test score, for example, to every one of these voxels and identify where the correlations are statistically significant. The location of any finding from any image analysis can be described with a system of standard spatial coordinates (height, width, depth) or by a standard nomenclature of brain areas differentiated by cellular structure, originally determined by Brodmann. Brodmann areas (BAs) are shown in Figure 3.6. Often, both BAs and spatial coordinates (usually based on the Talairach brain atlas or on the Montreal Neurological Institute’s system) are included in research reports.

  Figure 3.5 The VBM technique starts with an image (left) and automated algorithms then separate gray and white matter (middle) and then a value reflecting density is assigned to each voxel in the image. This value can be correlated to IQ, age, or other variables. (Courtesy Rex Jung.)

  Figure 3.6 Brodmann areas (BAs) are a standard way to label different brain regions based on early autopsy studies of neuron organization.

  Some of the first voxel-based MRI analyses of intelligence were reported in children. One research group obtained MRIs in 146 children (mean age 11.7, standard deviation 3.5) and reported a correlation of .30 between FSIQ and gray matter volume of the anterior cingulate gyrus (BA 32) (Wilke et al., 2003). Another group studied 40 children (mean age 14.9, standard deviation 2.6) and reported that gray matter volume was correlated to different parts of the cingulate (BAs 24, 31, 32) and to areas in the frontal lobes (BAs 9, 10, 11, 47) and in the parietal lobes (BAs 5, 7) (Frangou et al., 2004). More recent and advanced imaging studies of children will be reported in the next chapter, but these early studies in children were important for demonstrating the potential of imaging to elucidate relationships between brain development and IQ scores. One additional early study exemplifies this. In the largest and most representative sample of children studied up to that time with MRI, a research group (Shaw et al., 2006) introduced another method of image analysis that determined the thickness of the cortex. This was not VBM. Rather, it used numerous cortical landmarks and Euclidian geometry to calculate thickness at many points around the c
ortex. They had a sample of 307 normal children (mean age 13, standard deviation 4.5) who completed IQ tests and MRI scans on multiple occasions over time. Cortical thickness (CT) was correlated with IQ, but there was a clear developmental sequence showing a dynamic relationship between regional CT and intelligence as the brain matures through childhood and adolescence. The strongest correlations between IQ and CT were found in late childhood (approximately 8–12 years). These correlations were positive and they were found in areas throughout the brain. However, there was a difference between high- and average-IQ individuals. The high-IQ subjects showed “an initial accelerated and prolonged phase of cortical increase, which yields to equally vigorous cortical thinning by early adolescence.” This interesting finding, however, requires replication and there are new studies we will discuss in the next chapter. The Shaw et al. study, published in Nature, a prestigious science journal, and funded by the National Institute of Child Health and Development (NICHD), further added credence to investigations of intelligence/brain relationships.

  About the same time, VBM was applied in studies of intelligence in adults for the first time. We obtained MRIs for 47 adults across a wide age range (18–84) and correlated gray and white matter to IQ scores, correcting for age and sex (Haier et al., 2004). The results showed gray matter correlations in several areas distributed across the brain in all four lobes and in both hemispheres. One correlation between IQ and white matter was prominent in the parietal lobe (near BA 39). When we reanalyzed the data separately for males and females (Haier et al., 2005), we were surprised to see different results. In men the largest brain areas where more gray matter went with higher IQ were in posterior regions, especially in a part of the parietal lobe related to visual spatial processing. However, in women, almost all the areas where gray matter correlated to IQ were in the frontal lobes, especially around a part of the brain related to language called Broca’s Area.

  As with our previous PET study of mathematical reasoning, males and females showed different patterns of correlations. An unsettled issue is whether different male/female patterns are statistically significant. Nonetheless, these findings reaffirmed our view that all imaging studies of intelligence should analyze data separately for males and females just as routinely as groups of different ages are analyzed separately. Finding age and sex differences underscores one of our basic assumptions: not all brains work the same way.

  Even with the application of more standard VBM methods, however, many inconsistent results were reported. For example, one group (Lee et al., 2006) studied 30 older adults (mean age 61.1, standard deviation 5.18) and found only a correlation between Performance IQ and volume in the posterior lobe of the right cerebellum. Another group (Gong et al., 2005) studied 55 adults (mean age 40, standard deviation 12) and reported that correlations between gray matter and FSIQ were limited to areas in the anterior cingulate and the medial frontal lobes. As we have pointed out, often results were based on analyses that did not separate males and females, and restricted ranges may have limited correlations, as discussed in Chapter 1. Another key issue is the assessment of intelligence using IQ tests. Although the standard IQ tests provide a good estimate of the g-factor, IQ scores combine g and other specific intelligence factors. Would imaging results be more consistent if a better estimate of g was used?

  We addressed this in two studies (Colom et al., 2006a, 2006b) based on a reanalysis of our 2004 VBM data using the method of correlated vectors (Jensen, 1998). In this case, this method correlates the rank of g-loadings for each subtest of the WAIS to the rank of the same test correlation to gray matter. We found that g accounted for many of the FSIQ correlations with gray matter in the anterior cingulate (BA 24), frontal (BAs 8, 10, 11, 46, 47), parietal (BAs 7, 40), temporal (BAs 13, 20, 21, 37, 41), and occipital (BAs 17, 18, 19) cortices (Colom et al., 2006b). Moreover, in a separate analysis, we found a nearly perfect linear relationship between the g-loading of each subtest of the WAIS and the amount of gray matter correlated to each subtest score (Colom et al., 2006a). Thus, we come to another important observation. IQ tests have the advantages of a standardized test battery but the scores combine the general factor along with other specific factors. So the question of how intelligence correlates to brain structure and function depends on whether the question is about g or about more specific mental abilities. Inconsistent results among these early studies likely result from confusion on this issue as well as from issues about sampling and image analysis.

  3.8 Imaging White Matter Tracts with Two Methods

  A different kind of structural MRI is called diffusion tensor imaging, or DTI. Here, the MRI sequences are optimized to image the water content (i.e., hydrogen molecules) of white matter fibers and, when combined with special mathematical algorithms, the resulting images show white matter tracts in great detail. DTI measures can assess the density and organization of the tracts, which relate to how well they transmit signals. DTI is an excellent technique for identifying brain networks. Most DTI studies of intelligence are more recent and will be detailed in the next chapter. Here we note the first DTI study of intelligence by Schmithorst and colleagues (2005). They studied 47 children aged 5–18 years. After correcting for age and sex, the strongest correlations between IQ and the density/organization of white matter fibers were found in frontal and parietal/posterior areas. They noted these findings were consistent with the Wilke et al. study that had used VBM methods. As with the other early MRI studies, this one added to the excitement of using new imaging techniques to quantify brain/intelligence relationships.

  Whereas DTI can quantify white matter density and organization, MR spectroscopy (MRS) can make neurochemical determinations of white matter integrity, another measure of how well signals are transmitted through the fibers. For example, MRS can determine N-acetylaspartate (NAA), a marker of neuronal density and viability. The early MRS methods, however, were limited to single-voxel analysis, so the entire brain could not be studied at once. There were three early studies of intelligence using MRS. Jung’s research group studied 26 college students and placed the NAA measurement voxel in white matter underlying BA 39, 40 in the left parietal lobe (Jung et al., 1999b). They found a correlation between NAA and FSIQ of .52. They replicated and extended these findings in a new sample of 27 college students where the same region showed the NAA/IQ correlation and control regions in bilateral frontal lobes did not (Jung et al., 2005). They also showed that the NAA/IQ correlation was higher in the subsample of women. In the third MRS intelligence study, another group reported a sample of 62 adults in a wide age range (20–75 years). Scores on the high-g vocabulary subtest of the WAIS-R were correlated to NAA in voxels underlying left frontal BAs 10 and 46 (r = .53) and left BAs 24 and 32 (r = .56) in the anterior cingulate gyrus (Pfleiderer et al., 2004). All these early MRI studies of gray and white matter structure were exciting because they found correlations between various psychometric test scores of intelligence and quantifiable brain characteristics both in specific locations and in the connections among them. This increased optimism for the potential of discovering not only “where” in the brain was intelligence, but also “how” intelligence is related to brain function.

  3.9 Functional MRI (fMRI)

  Functional MRI uses scanning parameters that image aspects of hemoglobin in red blood cells because hemoglobin contains iron and iron molecules are quite sensitive to the magnetic fields used in MRI. A sequence of very rapid images is made, thousands per second. These are interpreted as showing blood flow in the brain. Those brain areas that are most active during a task (often compared to a no-task resting condition) have greater blood flow; less-active areas have reduced blood flow. Whereas glucose PET scans show the accumulation of brain activity over 32 minutes, fMRI scans show activity changes almost second by second. fMRI is now the most widely used imaging technology in cognitive psychology research.

  The first intelligence study using fMRI was from a group at Stanford University (Prabhakaran et al., 1997). They used
individual items from the Raven’s test chosen for three different types of reasoning required to solve the problem. They found blood flow increased in frontal and parietal brain areas while seven young adults (age 23–30 years) solved each item. They did not look for correlations between amount of activation and task performance because only a few problems were used and each person answered all items correctly. By design, this eliminates individual differences in task performance. This approach is typical in many cognitive studies where any individual differences among subjects related to intelligence are essentially ignored. More recently, there have been excellent discussions of the individual differences approach and its potential for advancing cognitive imaging studies (Kanai & Rees, 2011; Parasuraman & Jiang, 2012).

  Even though fMRI had been used in hundreds of cognitive studies by 2006, only 17 studies included any measure of intelligence or reasoning. Of these 17 fMRI studies, all but three had sample sizes of 16 or fewer and there were a variety of control tasks (or a lack of any control task in some studies) and a variety of intelligence/reasoning measures. None of the measures in these early studies were based on a battery of tests to estimate the g-factor. Some of the tests used in these studies during the imaging included working memory (Gray et al., 2003), chess (Atherton et al., 2003), analogies (Geake & Hansen, 2005; Luo et al., 2003), visual reasoning (Lee et al., 2006), deductive or inductive reasoning (Fangmeier et al., 2006; Goel & Dolan, 2004), and verb generation (Schmithorst & Holland, 2006). This last study was unique for its impressive sample size of 323 children (mean age 11.8 years, standard deviation 3.7). Given all these various findings and methods in the early studies, could any consistent threads be identified?